Intel Xe Exporter

A lightweight Prometheus exporter for Intel Xe (Arc) GPU metrics. Reads /proc/*/fdinfo/ to expose per-process and aggregate GPU memory and engine utilization as Prometheus gauges.

Runs as a DaemonSet on Kubernetes, one replica per node with an Intel Arc GPU.

Metrics that are already made available by node-exporter (hwmon...) such as temperature, fan speed and power usage are not duplicated by this exporter.

Metrics

Metric	Type	Labels
intel_xe_vram_resident_bytes	Gauge	—
intel_xe_system_resident_bytes	Gauge	—
intel_xe_process_vram_resident_bytes	Gauge
pid, process, card, tile
intel_xe_process_system_resident_bytes	Gauge	pid, process, card
intel_xe_engine_cycles_total	Gauge	engine
intel_xe_engine_active_cycles_total	Gauge	engine
intel_xe_engine_capacity	Gauge	engine
intel_xe_engine_busy_ratio	Gauge	engine

busy_ratio is computed from counter deltas between polling intervals (5s). The first scrape after startup returns no utilization data.

Running locally

Requires root to read /proc/*/fdinfo/.

pip install -r src/requirements.txt
sudo python src/intel_xe_exporter.py

The metrics endpoint is available at http://localhost:9830/.

Kubernetes deployment

Prerequisites

• monitoring namespace exists
• Nodes run a Linux kernel with the intel_xe driver

Deploy

kubectl apply -f kubernetes/

This creates:

• DaemonSet — runs on every node, requires hostPID: true
• Headless Service — discovers pods for scraping by Prometheus

Image: ghcr.io/mnlipp/intel_xe_exporter:latest

Security

To read GPU metrics from other processes, container needs access to /proc/*/fdinfo/. This requires two things:

• hostPID: true — the container must see host PIDs. Without this, /proc only contains the container's own processes.
• CAP_SYS_PTRACE — the kernel blocks reading /proc/<pid>/fdinfo/ for processes that don't belong to the caller. This capability lifts that restriction.

All other capabilities are dropped (drop: ALL) to keep the surface area minimal. The container runs as root because /proc/*/fdinfo/ is readable by root only.

Testing

Deploy the manifests from above, then port-forward to a running pod:

kubectl port-forward svc/intel-xe-exporter 9830:9830 -n monitoring

In another terminal, fetch the metrics:

curl http://localhost:9830/

You should see Prometheus-formatted gauges for GPU memory and engine stats.

Building the image

buildah build -t ghcr.io/mnlipp/intel_xe_exporter:latest \
  -f image/Containerfile src/

The Dockerfile expects src/ as the build context, not the repository root.